This paper introduces a context-agnostic machine learning framework for intelligent fault detection and classification in medium-voltage (MV) distribution networks. A hybrid dataset was constructed by combining synthetic waveforms generated from an IEEE/CIGRÉ-compliant MV feeder model with publicly available fault records from open repositories such as Kaggle, IEEE Dataport, and the UCI Machine Learning Repository. Fault scenarios included single-line-to-ground, line-to-line, double-line-to-ground, and three-phase events under various resistance levels and fault occurrence times. Transient-sensitive features were extracted from three-phase current and voltage signals using Discrete Wavelet Transform (DWT) and complemented with two statistical descriptors, maximum absolute value and energy, selected for their computational efficiency in real-time settings, five supervised classifiers Random Forest (RF), Support Vector Machine (SVM), Multilayer Perceptron (MLP), Decision Tree (DT), and k-Nearest Neighbors (KNN) were then trained and evaluated using accuracy, macro F1-score, and cross-validation stability. Among them, RF achieved the best trade-off between accuracy and robustness, positioning it as a promising candidate for real-time integration into protection relays, SCADA systems, and intelligent electronic devices (IEDs). The proposed approach is scalable, and adaptable to diverse MV grid configurations worldwide, providing a robust foundation for smart grid protection and predictive maintenance strategies.